Redesigning source control at Anthropic
Anthropic's Scion shows agent-scale monorepos can outgrow git, but the supplied discussion questions the traffic figures and whether human review survives the shift.
Anthropic describes Scion, an in-house version control system built to replace git for its monorepo after AI-driven code generation pushed merges from about 2,500 to over 30,000 per day. Scion assumes always-online clients and a server that handles commits, merges, rebases, blame and history, while clients keep bounded caches and sparse or virtualized checkouts.
The company says the system cut infrastructure footprint 200x, sped checkouts from minutes to seconds and delivered a 25% CI CPU win, but it presents Scion as unsuitable for general use and relevant only to enormous monorepos.
The bits that matter
- Anthropic says merges grew from about 2,500 per day in May 2026 to over 30,000 by September, with main-branch history near 100 GB and 25 million daily checkouts.
- Scion replaces git's decentralized model with always-online clients and server-side commits, merges and rebases; clients keep bounded caches and sparse or virtualized working copies.
- Changesets are write-ahead-log entries storing changed files relative to the first parent plus content-addressed blobs; derived data such as blame, path history and skip lists is computed recursively.
- The architecture tiers scion-daemon, regional scion-cache and authoritative scion-servers; immutable content-addressed objects allow over 99% cache hit rates and no packfiles or repository maintenance.
- Anthropic reports a 200x smaller CPU and memory footprint than its prior git mirroring, checkouts in seconds rather than minutes, and a 25% CI CPU capacity gain.
- A seven-person team built Scion and the Antique review tool and migrated in five months; Anthropic says Scion is not a general git replacement and only makes sense for huge monorepos.
Discussion analysis
Analysis details
22 comments analyzed. Analyzed .
The sample selects active discussion branches and includes available parent comments. It may omit parts of the full thread. The analysis included 22 of 123 usable stored comments because of the input limit. Selected themes do not measure community opinion or how common a view is.
Source control remains necessary for agent-driven development
Replies to the proposal to abandon source control argue that agents still need branches, pull requests, tags/versioning, build reproducibility, and history to understand changes and detect revert loops. One commenter says agents use history more effectively than comments or separate documentation; another extends the thought experiment to agents writing directly to production or assembly, with the reply that assembly is still source code.
Scaling version control for huge repositories and high merge rates
One detailed comment describes Anthropic's constraints: no-cap token spend, very large agent swarms, a 100GB uncompressed/25GB compressed git history, roughly 30,000 merges per day, and networked agents that make full local pulls too slow. It says Scion draws on Google Piper and Meta Sapling and is an online, download-only-what-you-need system, with prior version-control tradeoffs misaligned with these scaling needs.
AI-generated code quality, review burden, and tech debt
Commenters debate whether fast AI code generation avoids or accepts janky code. Some report that skipping review leads to poor abstractions, lack of refactoring, and accumulating AI tech debt; others defend the tools' quality as comparable to or better than average commercial software and argue perfection is not the goal. A further point is that code quality and interface design pay off in large codebases, where LLMs make wrong compromises.
Business strategy: selling tokens versus building end-user services
One comment asks why Anthropic would sell tokens to other companies instead of using its agents to build highly profitable services itself, such as a Google search competitor. The reply uses a gold-rush analogy: sell shovels, because the AI is the shovel.
Accelerationism, overproduction, and societal costs
Comments question what high churn and seemingly infinite merge demand accomplish and what real problems the code solves. A reply criticizes the accelerationist mindset, pointing to overproduction of lines of code, data-center and ecological impact, and asking whether software quality, government services, food, the environment, or working hours are improving.
Fast code churn as a defensive security strategy
One comment speculates about an 'Evolutionary Architecture' in which agents change and reimplement code so quickly that attackers probe code that no longer exists in production. The comment frames this as a possible defensive side effect of rapid change, even if security is not the reason for the change.